New usages for LLMs
In addition to innovations in the training and execution of LLMs, recent work has also focused on new usages of these models and improvements in their existing capabilities. A fundamental challenge of those capabilities is the propensity for LLMs to exhibit inaccurate, hallucinated output. We start this section by discussing some recent advancements in mitigating hallucination, before turning to novel applications in multi-modal and agentic models.
Detecting hallucinations
A core challenge of LLMs is that their primary objective is to generate tokens, not necessarily to produce factually accurate representations. This capacity to create outputs that seem plausible but are inaccurate is known as hallucination.1 Such hallucinations can either be factually inaccurate or inconsistent14 (Figure 10.7). Factual hallucination refers to the model creating incorrect information, while faithfulness hallucination refers to creating content inconsistent with the user...
